1// Copyright 2020 Google LLC 2// 3// This source code is licensed under the BSD-style license found in the 4// LICENSE file in the root directory of this source tree. 5 6$assert BATCH_TILE % 4 == 0 7$assert BATCH_TILE >= 4 8$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ" 9#include <assert.h> 10 11#include <wasm_simd128.h> 12 13#include <xnnpack/vunary.h> 14#include <xnnpack/common.h> 15 16 17void xnn_f32_relu_ukernel__wasmsimd_x${BATCH_TILE}( 18 size_t n, 19 const float* x, 20 float* y, 21 const union xnn_f32_relu_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN 22{ 23 assert(n != 0); 24 assert(n % sizeof(float) == 0); 25 assert(x != NULL); 26 assert(y != NULL); 27 28 const v128_t vzero = wasm_f32x4_splat(0.0f); 29 30 $if BATCH_TILE > 4: 31 for (; n >= ${BATCH_TILE} * sizeof(float); n -= ${BATCH_TILE} * sizeof(float)) { 32 v128_t vacc${ABC[0:4]} = wasm_v128_load(x); 33 $for N in range(4, BATCH_TILE, 4): 34 v128_t vacc${ABC[N:N+4]} = wasm_v128_load(x + ${N}); 35 x += ${BATCH_TILE}; 36 37 $for N in range(0, BATCH_TILE, 4): 38 vacc${ABC[N:N+4]} = wasm_i32x4_max(vacc${ABC[N:N+4]}, vzero); 39 40 wasm_v128_store(y, vacc${ABC[0:4]}); 41 $for N in range(4, BATCH_TILE, 4): 42 wasm_v128_store(y + ${N}, vacc${ABC[N:N+4]}); 43 y += ${BATCH_TILE}; 44 } 45 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) { 46 v128_t vacc = wasm_v128_load(x); 47 x += 4; 48 49 vacc = wasm_i32x4_max(vacc, vzero); 50 51 wasm_v128_store(y, vacc); 52 y += 4; 53 } 54 if XNN_UNLIKELY(n != 0) { 55 v128_t vacc = wasm_v128_load(x); 56 57 vacc = wasm_i32x4_max(vacc, vzero); 58 59 if (n & (2 * sizeof(float))) { 60 *((double*) y) = wasm_f64x2_extract_lane(vacc, 0); 61 vacc = wasm_v32x4_shuffle(vacc, vacc, 2, 3, 2, 3); 62 y += 2; 63 } 64 if (n & (1 * sizeof(float))) { 65 *y = wasm_f32x4_extract_lane(vacc, 0); 66 } 67 } 68} 69